TLDR AI 2025-05-05

Apple is reportedly working with Anthropic to build a next-gen internal coding assistant using Claude Sonnet, integrated into Xcode to generate, edit, and test code.

After reports of overly agreeable behavior in GPT-4o, OpenAI has detailed plans for more transparent updates, a new opt-in alpha phase, and stricter behavior safety reviews.

The new synthetically trained reasoning model from Microsoft shows strong local math and code performance, even in the face of poor internal world knowledge.

OpenAI's release of GPT-4o resulted in extreme sycophancy and other issues, raising concerns about its detachment from the OpenAI Model Spec against flattery. This was likely a misstep in pursuit of user engagement, exacerbated by A/B testing favoring positive responses. OpenAI CEO Sam Altman acknowledged the problem and has promised fixes. The situation highlights the risks of models being optimized in ways that might betray user trust.

MCP's main functions are advertising and calling tools, similar to OpenAPI, but it offers a more streamlined approach. While both can achieve similar outcomes, MCP's simplicity and reduced size make it attractive. MCP's perceived necessity is more sociological than technological.

FUSED introduces sparse unlearning adapters to selectively remove knowledge in federated learning, making unlearning reversible and cost-efficient.

Google Research has introduced a global data collection initiative co-created with local experts to improve AI relevance in underserved regions.

WebThinker is a deep research framework fully powered by large reasoning models (LRMs). It enables LRMs to autonomously search, deeply explore web pages, and draft research reports.

Alibaba has released Qwen 3, a family of AI models they claim can rival top models from Google and OpenAI. These open-licensed models feature advanced reasoning capabilities and a mixture of expert architecture. Supporting 119 languages and trained on 36 trillion tokens, Qwen-3-235B-A22B notably outperformed OpenAI's o3-mini in several benchmarks. It's not yet publicly available.

This method enhances image generation by leveraging self-attention features from pretrained diffusion models, introducing an attention distillation loss to optimize stylization and accelerate synthesis.

Google's SpeciesNet is an open-source AI model for identifying animal species from camera trap images. The model, previously used in Wildlife Insights, will help scale biodiversity monitoring efforts.

A livestream featuring Google's Gemini AI has completed Pokémon Blue, drawing praise from Google executives despite being an unofficial effort.

Forerunner, known for investments in Oura and Chime, anticipates future growth in the AI era.

OpenAI has rolled out improvements to ChatGPT's search and added a streamlined shopping experience with product details, reviews, and buy links.